4.3. Verification of Extraction Results by Different Methods
Due to the abnormal value of the image pixels, the water body extraction result inevitably has a small range of noise. Therefore, it was necessary to further process the extraction results to get the final result. Firstly, the water extraction results were converted from the raster format to the vector format, and then the Aggregate Polygon operation was carried out in ArcgisPro2.5. The reason for executing the Aggregate Polygon operation was that in the conversion process of water extraction results, some water bodies were divided into several small water bodies, which needed to be aggregated into one water element. We set the Aggregation Distance to 60 m, and then multiple water elements with a distance of less than 60 m were aggregated into one water element. Due to the limitation of image resolution, it was difficult to identify particularly small water bodies (<100 m2), so these water bodies were not included in the statistics. In the aggregation operation, Min Area and Min Hole Size were set to 100 m2, bodies of water less than 100 m2 in area were removed, and holes less than 100 m2 between vector bodies of water were filled as bodies of water. In order to compare the effects of different methods for extracting small water bodies, four regions (small rivers, aquaculture water, small urban reservoirs and small ponds in mountainous areas) were selected for comparative analysis.
Because of image resolution limitations, small rivers show small and narrow spatial characteristics on the image. Mixed pixels interfere with the extraction process of small rivers, resulting in discontinuity of the extracted river water. As shown in Figure 10
, the RF method was not as good as the SVM method for the classification of mixed pixels, and the water body was missing in the red frame, while the SVM extraction result was relatively complete. The
method misidentified many non-water bodies pixels, and the main ones were terrain shadows.
had good results in removing topographic factors, but the identification of water bodies was incomplete. The main reason is that the
method can remove shadows, which is suitable for areas where shadows are the main noise, and
method can effectively eliminate non-water pixels on dark building surfaces and is suitable for scenes where shadows are not the main noise [15
]. The water body information extracted by MNDWI is also mixed with a lot of shadow noise.
The area of aquaculture water is small, and its shape is generally rectangular. The nutrient level of aquaculture water has a certain impact on the spectral characteristics of a water body. There are generally pond ridges between different aquaculture water surfaces. The SVM method was not complete in extracting water bodies. The main missing locations were mostly located at the edge of the pond, which were mainly mixed pixels of water and non-water bodies (Figure 11
). The water body extracted by SVM was complete, but the error rate of extraction rate was high. Not only the ridge of the pond but also the road next to the pond (inside the red frame) was identified as a body of water. In the red box of MNDWI extraction results, there were two slender aquaculture ponds with a pond boundary between the ponds. The MNDWI method identified the pond boundary as a water body and two small ponds as large ponds. In the red box, the body of water in the pond was not visible. In the identification of water bodies, some pond ridges were identified as water bodies, and the boundary between ponds was blurred. In contrast, the NDWI method and the MFTSA method had no missing extraction, and the water surface boundaries of different ponds were clear.
In urban areas, artificial building information is the dominant background information. The reflection of buildings is strong, while the reflection of water bodies is weak (Figure 12
). Therefore, large background buildings and building shadows strongly interfere with water body extraction. The SVM method extracts water bodies incompletely. SVM seldom misses water bodies’ pixels in other scenes, but it misses water bodies in small urban reservoirs. The main reason is that small water bodies in cities appear black, and the spectral characteristics of water bodies in other places are not consistent. NDWI, MNDWI,
are relatively complete in extracting water bodies, but there are varying degrees of false extractions. Therefore, the RF and MFTSA methods are better in the identification of small urban water bodies.
There are hill shadows between mountains. Although we used slope information to remove the hill shadows, it could only remove the hill shadows on the larger slopes. In some mountainous areas with smaller slopes, the effect of removing hill shadows was not good enough. As shown in Figure 13
, the extraction results of the RF method leaked a part of the water body in the red frame, resulting in holes in the pond; SVM, MNDWI and
recognized some mountain shadows as water bodies, and
had the most serious misrecognition phenomenon, mainly because the index was suited for scenes without shadows.
eliminates shadow pixels that are easy to confuse with water bodies in the result of
, so the effect of extracting water bodies is better. On the whole, the NDWI,
and MFTSA methods were better in extracting small water bodies in mountainous areas.